I love the idea of all AI grading. In fact, I have no question that it will takeover the entire market eventually. Is this company going to be the one? Maybe. They are first and are being run by a very impressive team. Plus, they are well funded.
After a couple of us met with them back in February I was extremely impressed. I was considered an honorary “founder”
I needed to test them out though and am now ready to do so.
I submitted an already graded and cracked PSA lot of the below 10 cards. I issued them a “Gary’s Opinion” pregrade. Then I got the PSA grades after a year. Now I’m submitting them to AGS to get their grade. I will show all that info when they return.
Finally, I’ll crack them and have them regraded by them in person at Collect-A-Con Orlando in February.
After that we can all compare the results which should prove interesting and informative.
Here’s the 10 cards. They are all in pack pulled condition.
My name is Alex and I am the CEO of AGS (Automated Grading Systems). Our AI grading service is called RoboGrading and is only available for Pokemon cards… for now. We officially launched at Collect-A-Con on Oct 16th and graded cards live with our system on the spot. Since then we have graded 3,500+ cards and are growing quickly.
My team and I started AGS out of a passion for collecting and lack of good options currently available for collectors. We’ve built a few successful tech companies before and know that AI grading can fix the current issues the industry is facing.
I wanted to start this conversation to answer any questions anyone has and maybe offer some an opportunity to try it out. It starts at only $20/card but maybe I can allow a few people to grade for free (DM me, limited availability)
Looking forward to bringing more transparency and awareness.
How well can the AI assess surface issues? For instance, if a card has an nearly imperceptible indent – one that a human grader would notice, but that doesn’t deform the card in a significant way.
Where a camera or pure computer vision can fall short, our laser is easily able to pick up things like that. Each micron of the card we scan gives us a height value which can be analyzed by our AI.
How do you intend to win a market share by effectively promising lower net grades (due to the alleged precision involved) and by charging $250-$2000 per high-value card as a startup company with absolutely zero street cred?
How do you intend to get the data required for us to assess the precision of your process? Let’s say for the sake of argument that I’m going to send cards worth 50K+ to you (which I doubt I will given the ridiculous $2000 per card price tag): how would I know if your grades can be trusted? I’m sure you can drum up a decent sample size for modern stuff and light played unlimited wotc in no time, but what about the good stuff in gem mint/mint condition? Is something in the works, do you have the backing of some whale or do you hope to build the sample size over time?
Your website is filled with “our A.I can do this and that”, but no in depth explanations as to your process. I’m not saying you should spill the secret sauce, but you’ll have to reveal some more details if you mean to convince people about the process itself. “Because A.I” doesn’t cut it here.
The 1st ed Charizard you graded in the Youtube video received a grade of 8, but apparently the surface was a 6 (and the corners a 10.) As someone who, in accordance with most of the other grading companies, puts a premium on surface, this discrepancy would concern me. Could you go into further details about how the sub-grades affect the overall result?
I think machine grading is the way of the future. What would make AGS more compelling are maybe some case studies on the website demonstrating that your automated solution is superior (to which end you would have to define) to human grading? I’m no computer vision expert, but I’m skeptical that any automated solution is of a high enough fidelity to compete with some of the more established grading companies at present without a manual review step (PSA/BGS have their moments with inaccuracies, but these seem like the exceptions rather than the rule – a consequence of human error, which you guys are no doubt trying to address)
Another question: Off your website – “precision” and “easy” seem to be at odds with each other, no? I imagine the value that you guys are trying to bring to the table is being able to: 1. identify all imperfections on a card, and 2. roll said imperfections up into a numeric score. Not sure what the purpose of the app is here unless you’re targeting a segment of users that are just looking for a high-level assessment of their card (IMO the descriptions of each level of the 10 point grading scale should be more than sufficient for this).
Your “AI grading” refers to your imaging and recognition of card flaws, correct? The grading itself is based on human input or some set of rules/factors that your team determines?
“Micron” being hyperbole here, right? When your resolution is that fine, I’m sure the entire surface of the card might look like some sort of craggly mountain with micro scratches and manufacturing imperfections too small for even a human with a loupe to pick up.
“I’m not saying you should spill the secret sauce”
^
There should definitely be transparency into their process and how grades are determined. These guys are essentially an imaging company – their secret sauce is in how they scan (visually or otherwise) and process card imperfections, not in how they construct a grading scale (forming a set of rules to assign grades is trivial in comparison).
Detection of imperfections is also a non-trivial task. A case study into how AGS operates would be great. Hand wavey “AI, machine learning” doesn’t quite cut it in 2021… I think the technical problem at hand is much more complex than the surface level “cameras/lasers detect flaws → AI → grade” flow some might envision
Related to this, the textures of holos from certain sets (i.e., Expedition, Aquapolis) are often naturally “wavy.” That sort of wavy texture, though, could be actual damage on holos from other sets. If the AI is precise enough to detect any sort of surface abnormalities, it will have to take into account what set a card is from. Condition can’t just be evaluated in a vacuum.
@wooter, I’d also be curious to know if the AI differentiates between holo scratches and factory print lines. Basically, can the AI differentiate between factory flaws and handling flaws? And, if so, do they each affect the grade to the same extent?
Maybe not funding, but I can see a group of reasonably smart, competent folks coming up with a solution that beats out PSA/BGS with respect to grading consistency. Whether or not that’s enough to take a meaningful slice of the pie is something else.
One thing is for sure: their pricing (especially as a new entrant with no reputation or indication that their solution *is* better) is disappointing. There are loads of new grading companies that have cropped up in the past few years — IMO, touting “AI” and showing off some imaging hardware is hardly enough to differentiate yourself in that environment.
I don’t think we are promising lower grades. Cards that deserve 10s and are gem mint, get them. We are promising more consistency in the grades along with full transparency on why each card got the grade it received.
The more cards we grade the better our AI becomes over time. We do have some key members on our team and are currently speaking with experienced veterans to further enhance our credibility.
We actually show many videos on our Youtube of our process and have shown it live at Collect A Con from A-Z. The secret sauce is the patent pending machine that scans the card and our AI models which are constantly updated.
Zack, aka Gem Mint Pokemon let us grade this card very early on in June, before we launched. There is 4 subgrades on front and 4 on the back. That averages out to give you the total score.
We have paused our app scoring feature and are using the mobile app to accept submissions faster. Our app was intended to give a soft score and make it easier to trade in person or online. So you can take a photo of your card, get a soft grade to get an idea of what it would grade. Would be huge to get these grades on ebay listings… Surface is really hard to grade from a photo. Also the lighting and the angle affects the soft grade our app gave. So we decided to take down this soft grading for now as it wasn’t accurate. We will release an upgraded app in due time but our current focus has been our paid grading service: Robograding.
Our AI grading consists of two parts: annotating all of the defects and assessing a score to each defect.
In what way does the AI get better? I assume you’re talking about detection of imperfections as you add more and more scans to your internal repository of scans, and not the grade, which relies on some combination of count and severity of these imperfections? If so, can you comment on the current level of fidelity?
The videos on YouTube don’t show much at all… and “patent pending machine” and “AI models” are incredibly reductive. Any detail on this?
The annotations do detect scratches and print lines differently because of the heights. The grade would be affected depending on size and depth of the flaw.
That’s awesome. For any technical folk (as well as the lay person), I think case studies or documentation (without revealing your secret sauce, yes) will go a long way towards establishing credibility and trust in the process.
Would be interesting to see an engineering blog emerge – this is quite an interesting problem that I’m sure can be discussed and shared at a high level without revealing proprietary details (see: Stitch Fix’s engineering blog for example, where they provide a fantastic top level discussion on their data science approach and algorithms)
Each scan gives us 12 million data points on the card. The CSV file before compression is 120mb.